Andreas Christe, Presenter: Nothing to Disclose

Marios Anthimopoulos, Abstract Co-Author: Nothing to Disclose

Stergios Christodoulidis, Abstract Co-Author: Nothing to Disclose

Stavroula Mougiakakou, Abstract Co-Author: Nothing to Disclose

The classification of HRCT image patches with interstitial lung disease (ILD) abnormalities, as a basic component towards the quantification of the various ILD patterns in the lung.

Based on the publicly available TALISMAN database consisting of 113 HRCT scans, a dataset with nearly 2500 ILD image patches was created with size equal to 21×21pixels. Six lung patterns were considered: normal, ground glass opacity (GGO), consolidation, reticulation, honeycombing and the combination of reticulation with GGO. Initially each patch is described by a feature vector which is then fed to a machine learning classifier. Feature extraction relies on a filter bank containing the 25 basis functions of the 5x5 Discrete Cosine Transform (DCT). After convolving the image with the filter bank, the 10-quantiles are computed on the filter responses for describing the distribution of local frequencies that characterize image texture. Quantiles are points taken at regular intervals from the cumulative histogram of the image; 10-quantiles are 9 values splitting the histogram to 10 intervals. Moreover, the minimum and maximum value of every filter response is added, together with the 32 gray-level histogram values of the original image. The final feature vector with 307 values is fed to a random forest (RF) with 40 trees for the classification.

The proposed ILD pattern methodology achieved an overall accuracy in the order of 90% outperforming state-of-the-art methods tested in the same data, by at least 7%. The sensitivity (%)/specificity (%) were: normal - 98.8/97.8; GGO - 81.3/98.8; consolidation - 92.7/99.5; reticulation - 85.6/95.7; honeycombing - 86/98.9; and combined reticulation/GGO - 88.2/94.8.

The combination of the proposed DCT-based features with RF classification showed very promising results outperforming many state-of-the-art methods. Future work includes investigating of the extension of the proposed 2D fixed-scale filter bank to multiple scales and three dimensions.

DCT-based features and random forest classification are powerful tools from the fields of computer vision and machine learning which can help in the field of lung CT image analysis for the diagnosis of ILDs.

Christe, A, Anthimopoulos, M, Christodoulidis, S, Mougiakakou, S, Classification of Interstitial Lung Disease Patterns Based on Local Discrete Cosine Transform Features of HRCT Images.  Radiological Society of North America 2014 Scientific Assembly and Annual Meeting, - ,Chicago IL. http://archive.rsna.org/2014/14019468.html